What question did this study set out to answer?

To develop techniques for verifying the provenance of LoRA adapters and detecting unauthorized modifications.

June 3, 20260 citationsOpen Access

Merkle-Chained LoRA Genealogy and Hopfield Attractor Watermarking for Adapter Provenance Verification

WWWeslyn Cory WhiteheadKlamath Community College

Puntos clave

To develop techniques for verifying the provenance of LoRA adapters and detecting unauthorized modifications.
Introduced Hopfield Attractor Watermarking to embed patterns as LoRA component deltas.
Implemented Merkle-Chained Adapter Genealogy to create a verifiable tree structure using SHA-256 hashes.
Combined both techniques to provide a unified verification API.
Successfully detected adversarial weight modifications in LoRA adapters.
Prevented unauthorized derivations and base model spoofing effectively.
Showed compatibility with Q4_K_M quantization through 8-bit fingerprinting.

Resumen

We present two complementary techniques for LoRA adapter provenance. Hopfield Attractor Watermarking (HAW) embeds Hopfield-style attractor patterns as rank-1 LoRA component deltas, verifiable via a sentinel forward pass and tolerant of Q4KM quantization via 8-bit fingerprinting. Merkle-Chained Adapter Genealogy (MCAG) chains adapters into a verifiable Merkle tree via SHA-256 parent hashes. Together they detect adversarial weight modification, unauthorized derivation, and base model spoofing. Applied to the LoRA adapter proliferation problem on public model hubs. The combination addresses three practical attack vectors with a unified verification API and no training-time modifications.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo

Cite This Study

Weslyn Cory Whitehead (Mon,) studied this question.

synapsesocial.com/papers/6a1fc5d7dee9eb8c0dce7327 https://doi.org/https://doi.org/10.5281/zenodo.20484511

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo